Fig 1.
Art analysis using Levinson’s definition of art.
(a) Art consists of Exhibited Properties (EXPs) and None-Exhibited Properties (NEXP). (b) Art understanding is gained by relating EXPs to NEXPs rather than merely looking at EXPs. (c) The difficulty in art understanding is shown as a spectrum with an example for each end: top, “Fountain” by Marcel Duchamp, a conceptual piece with a greater significance of NEXPs, and bottom: “The Accident,” by William Geet, a figurative piece with a more literal visual narrative and therefore more significance of EXPs.
Fig 2.
Computational art understanding tasks.
NEXPs are more critical as more artistic aspects must be considered (i.e. fewer NEXPs are needed for style recognition and more NEXPs for gallery recognition).
Fig 3.
(a) The methodology to validate hypotheses I and II.(b) Dataset design method.
Fig 4.
The designed datasets S1, S2, S3, G1, and S4 pertain to a broad difficulty spectrum due to their EXPs and NEXPs dissimilarities and similarities within or across galleries.
Fig 5.
Individual examples from dataset S1.
(a) “Translucent Hat” from Gallery “Heat+ High Fashion”. (b) “Hairpiece” from Gallery “Mukono”. The two photos illustrate EXP similarities between the galleries of the dataset. Black and white photography and the high contrast of a dark figure on a light background are some EXPs that the two photos and their respective galleries have in common. These similarities between the galleries of a dataset increase the classification difficulty.
Fig 6.
Gallery outliers for dataset S1.
Galleries Heat + High Fashion, My Mother’s Clothes, and Scene do not have any outliers.
Table 1.
Characteristics of dataset S1: The related galleries (e.g., names of the exhibitions or shows), the number of images per gallery, and the total number of images per dataset.
Table 2.
EXPs of the galleries in dataset S1.
Table 3.
NEXPs of the galleries in dataset S1.
Fig 7.
(a) PCA plots for the training and test data of subsets difficult, average, and easy. (b) Box and whisker plots of the overall accuracies (ACC) of the DCNN classification of the four subsets. Subsets difficult and average have outliers of 20% and 100%.
Fig 8.
Class-wise classification metrics results for dataset S1.
Fig 9.
Class-wise classification metrics results for dataset S2.
Fig 10.
Class-wise classification metrics results for dataset S3.
Fig 11.
Class-wise classification metrics results for dataset G1.
Fig 12.
Results for dataset S4 and “Non-Art”’s 34-image/18-image versions.
(a) Class-wise classification metrics results. (b) Box and whisker plots of the overall accuracies (ACC) of the DCNN classification for datasets S4–34 and S4–18.
Fig 13.
Gallery themes summary.